| Sample of Swing Metrics (2024) | |||||
|---|---|---|---|---|---|
| Batter | Bat Speed | Swing Length | Swing Path Tilt | Attack Angle | Attack Direction |
| Altuve, Jose | 70.3 | 8.3 | 29.9 | 17.3 | −5.3 |
| Judge, Aaron | 73.5 | 8.2 | 40.0 | 16.4 | −2.2 |
| Ohtani, Shohei | 72.6 | 7.2 | 31.1 | −1.9 | 1.9 |
Clustering Swing Behavior Within MLB Hitters
Introduction
In baseball, a pitcher controls the game more than the hitter. The pitcher can take his time to select the pitch type and location carefully. Once the ball leaves the pitcher’s hand, it is the batter’s job to discern the information in milliseconds. He has to decide if he wants to swing and where to swing, all in the time it takes for a person to blink. This is part of the reason hitting is so difficult, and why even the best players fail seven out of ten times.
Mirroring the game itself, baseball analytics have also favored the pitcher. Pitchers have gained access to their ball tracking data, which provides them with more information to tinker with and make them even more effective. Previously, batting metrics were also reactive; launch angle and exit velocity were two factors that hitters tried to optimize, but they were still heavily dependent on the pitch and their swing, for which they had no actual data. Recently, analytics have caught up to the offensive side of the game. Batters and analysts alike now have access to bat-tracking data. Five key metrics are bat speed, swing length, swing path tilt, attack angle, and attack direction. Now, hitters have access to data that they can use and study to optimize their mechanics and improve their effectiveness at the plate.
A lot of exploratory research has been conducted to identify swing archetypes among all Major League hitters. This research aims to apply a single approach to a group of hitters. We are attempting to go deeper and discover trends at the individual hitter level. For example, many hitters claim to have a “two-strike approach,” where they protect the plate and shorten up to put the ball in play, perhaps sacrificing power to reduce the risk of a strikeout. Or maybe hitters adopt a more “uppercut” swing when there is a runner on third and fewer than two outs, trying to hit a sacrifice fly.
The catalyst for this project is based on these assumptions. In the past, we could only watch the TV and guess what these hitters were doing. Now we can quantify it and examine their approaches under the hood. We aimed to investigate our assumptions about hitters’ approaches by clustering their swings into distinct groups and identifying when hitters used different swing types.
Data
All data used in this analysis is from the 2024 MLB season. This is because, given the newness of the data, it is the only full season of bat-tracking data we have. Each observation is a swing from the 2024 season, complete with pitch characteristics, bat-tracking, and the result of the play. We are using data from Baseball Savant, scraped using the sabRmetrics package. Here are some example rows:
Data Cleaning
For our analysis, we only wanted to consider full swings where the batter fully intended to make contact with the ball. We didn’t want to include check swings because the hitter is intentionally trying to stop himself from swinging and take the pitch. We also omitted bunts, as those are not true swings.
To aid in our analysis, we added two features to the data: squared up and blast. These are both indicators with values “TRUE” or “FALSE.” We created them using the definitions on Baseball Savant. A swing has “squared up” a ball when the exit velocity is at least 80% of its theoretical maximum. That percentage is also referred to as the squared-up rate. A blast is a swing where the squared-up percentage plus the bat speed equals at least 164. Adding these features enables us to evaluate the quality of a player’s swing, disregarding the outcome in the field.
Methods
Because our goal was to identify distinct swing types for each hitter, we needed a method to evaluate the similarity between swings and to determine natural groupings within that space. To accomplish this, we applied Gaussian mixture modeling (GMM) to cluster each player’s swings based on their swing-tracking characteristics. We fit each hitter’s swing to a GMM by using R’s mclust package. The model was fit using an Expectation-Maximization algorithm. Each swing was defined by the following features: bat speed, swing length, swing path tilt, attack angle, and attack direction. We chose to include these features, as they are each core components of a player’s swing.
We selected GMM over hard clustering approaches like k-means for two key reasons:
Flexibility: GMM assumes the data arises from a mixture of Gaussian distributions, allowing it to model soft cluster boundaries and capture overlapping clusters. This is particularly important given that hitters may exhibit multiple swing “types” depending on context (e.g., adopting a more protective approach in two-strike counts).
Adaptive model complexity: Rather than manually specifying the number of clusters for each player, we used the Bayesian Information Criterion (BIC) to select the optimal number of components automatically. This approach scales well across the league and avoids imposing a rigid or uniform structure across all hitters.
These advantages were critical for a project of this scope, where we clustered swings individually for every hitter in MLB. Using a fixed number of clusters would have ignored inter-player variability, while manual tuning would be labor-intensive and introduce potential subjectivity or bias.
Although we briefly considered applying Principal Component Analysis (PCA) to reduce dimensionality before clustering, we found that doing so did not substantially alter the clusters or resulting visualization. Ultimately, we retained the raw swing features to preserve interpretability.
Results
After applying GMM clustering to each MLB hitter’s swing data, we used mosaic plots to compare how the distribution of swings changed in certain situations. One situation we examined was how a player’s swing changed when they had two strikes against them. When a hitter is one pitch away from striking out, they may change their swing in an attempt to put the ball in play and make something happen. Hitters may claim to “shorten up” or choke up on the bat, but now we can look behind the curtain and see what is going on.
Looking at a few of these plots, a notable trend emerged: hitters tend to use “worse” swings in two-strike counts compared to other situations. Below you can see how contact and quality of contact vary by cluster.
| Altuve Contact Quality by Cluster | ||||
|---|---|---|---|---|
| Cluster | Number of Swings | Contact % | Squared Up % | Blast % |
| 1 | 337 | 69.73% | 22.85% | 2.67% |
| 2 | 114 | 74.56% | 14.04% | 2.63% |
| 3 | 629 | 85.21% | 30.52% | 9.86% |
| Albies Contact Quality by Cluster | ||||
|---|---|---|---|---|
| Cluster | Number of Swings | Contact % | Squared Up % | Blast % |
| 1 | 588 | 83.50% | 27.38% | 7.14% |
| 2 | 237 | 78.48% | 24.89% | 3.38% |
| Arenado Contact Quality by Cluster | ||||
|---|---|---|---|---|
| Cluster | Number of Swings | Contact % | Squared Up % | Blast % |
| 1 | 576 | 91.84% | 32.29% | 11.63% |
| 2 | 210 | 67.62% | 17.14% | 0.48% |
| 3 | 268 | 87.31% | 23.13% | 3.73% |
These findings challenge the common belief that players adopt a more disciplined, contact-oriented approach with two strikes. Importantly, by “worse” swing, we do not imply that hitters are consciously choosing less effective swings. Instead, contextual factors (pitch type, location, etc) appear to induce swings from the lower-performing regions of a player’s swing distribution. Our clustering doesn’t suggest that players have five entirely distinct swing mechanics. Instead, we’ve captured different areas within each player’s overall swing distribution. This allows us to examine when different swing types emerge and begin to explore why these shifts occur. A few such possibilities are the pitch location and pitch type. Here are how the swings are distributed when looking at those characteristics.
To help understand what we are referring to when we say “Cluster 1 swing” vs a “Cluster 3 swing”, below are two videos of Jose Altuve, both in two-strike counts.
